Outdated answers
Old policies and superseded document versions remain in the source collection.
Knowledge Readiness Sprint
Prepare your knowledge base for reliable AI answers.
We clean, structure, and validate the knowledge behind your AI assistant—addressing duplicate content, conflicting versions, and missing context in a focused three-week sprint.
From US$3,500 · Three-week sprint · Fixed scope before kickoff
Delivered by Richard Gregorio — AI technical architect with 20+ years in enterprise technology.
Is your source content holding your AI assistant back?
Conflicting sources
Different articles provide incompatible instructions, without a clear authoritative version.
Missing context
Useful information exists, but its structure makes relevant passages difficult to retrieve or interpret.
From conflicting sources to clear, traceable knowledge.
Synthetic demonstration. No customer data. A simplified excerpt showing one issue within a larger knowledge collection.
A support assistant needs to answer:
“I bought my first annual subscription 20 days ago. Can I get a refund?”
Before
Three sources. Conflicting guidance.
- Old help article: refunds within 30 days.
- Current policy: first-time annual subscriptions refundable within 14 days.
- Support playbook: requests at 15–30 days may receive discretionary Billing review.
The assistant may retrieve an outdated rule or miss the exception procedure.
After
One approved rule, with the exception connected.
- Standard refund window: 14 days.
- Days 15–30: refer to Billing for discretionary review.
- Approval is not guaranteed.
- Source references and policy-owner approval are recorded.
The prepared entry keeps eligibility, exceptions, and next steps together.
-
Source register
Where the rule came from.
-
Conflict log
What required an owner’s decision.
-
Validation checklist
How the prepared knowledge is checked.
Explore the preparation and validation details
Before: three sources, conflicting guidance
| Source | Content |
|---|---|
| Old help article | “Subscriptions are refundable within 30 days.” |
| Current billing policy | “First-time annual subscriptions are refundable within 14 days. Monthly subscriptions are non-refundable.” |
| Support playbook | “For annual purchases made 15–30 days ago, refer the request to Billing for discretionary review. Do not promise a refund.” |
The old article has no retirement label. The current policy and exception procedure sit in separate documents.
The knowledge problem: retrieval may surface the outdated 30-day rule or miss the distinction between refund eligibility and discretionary review.
What we change
- Flag the conflicting refund windows for the designated policy owner.
- After approval, retire the outdated article from the assistant’s active collection.
- Connect the current policy with its applicable exception procedure.
- Preserve source references, effective dates, and ownership.
- Structure the content so eligibility, exceptions, and next steps remain together.
After: one traceable policy entry
- Topic
- Subscription refunds
- Applies to
- First-time annual subscriptions
- Standard eligibility
- Within 14 days of purchase
- Exception process
- Requests made 15–30 days after purchase may be referred to Billing for discretionary review
- Restriction
- Referral does not guarantee approval
- Sources
- Current billing policy and support playbook
- Authority
- Confirmed by the designated policy owner in this fictional scenario
What a correct answer should contain
For the question “I bought my first annual subscription 20 days ago. Can I get a refund?”
“For a first-time annual subscription purchased 20 days ago, the standard 14-day refund window has passed. Your request can be referred to Billing for discretionary review, but approval is not guaranteed.”
This is an expected answer for validation, not an observed model output.
How we would validate it
| Check | Expected behavior |
|---|---|
| First-time annual purchase, day 10 | Explain standard refund eligibility |
| First-time annual purchase, day 20 | Explain discretionary review without promising approval |
| Monthly subscription, day 10 | State that monthly subscriptions are non-refundable |
| Subscription type unspecified | Ask which subscription the customer purchased |
| Source traceability | Reference the approved policy and applicable procedure |
Retrieval checks run in an agreed environment. When that environment is a reference setup, the results do not establish how the assistant will perform in production.
What you receive
Knowledge Readiness Sprint
-
Prepared knowledge collection
Content cleaned and structured for the agreed assistant use case.
-
Source and version register
Traceable sources and versions, with authority confirmed by your team.
-
Conflict and gap log
Unresolved contradictions and missing information requiring an owner’s decision.
-
Validation results
Agreed content and retrieval checks, with findings and limitations documented.
-
Handoff package
The agreed export, preparation rules, and maintenance guidance.
Typical handoff
- Prepared chunks as JSONL with a manifest describing fields, sources, and versions.
- Or a folder of Markdown files with YAML front matter for titles, sources, and effective dates.
- A short written guide for loading or syncing into your agreed retrieval environment—your team runs production ingestion unless separately scoped.
From US$3,500
Three-week delivery window
Fixed scope agreed before kickoff
One assistant use case, one language, and one agreed collection of text-based content. Volume and complexity are confirmed after sample review.
Handoff uses the export format your team agrees at kickoff—typically structured Markdown or JSON chunks with retrieval metadata—with schema, file layout, and import steps documented in the handoff package.
Delivery starts after scope approval and receipt of required inputs and access. Outstanding content decisions may affect completion.
Scope details and separately quoted work
Your team approves authoritative sources and resolves conflicting business policies.
Delivery format
- Export format, encoding, and chunking rules are fixed in the sprint scope before preparation begins.
- Deliverables include the prepared files plus a short guide for loading or syncing them into your agreed retrieval environment.
- Automated production ingestion, custom connectors, and platform-specific deployment are separately scoped.
Examples we align to at kickoff—not a fixed product list—include Salesforce Data Cloud and Knowledge, Azure AI Search, and comparable enterprise retrieval stacks. The export is chosen for your platform during scope approval.
Separately quoted
- Custom connectors and integrations.
- Extensive OCR or complex extraction.
- New content creation.
- Production deployment.
- Ongoing maintenance.
- Broader governance or permissions redesign.
How the sprint runs
-
1. Confirm fit and scope
We review your use case, representative source samples, and target outcome, then agree on scope, deliverables, and acceptance checks.
-
2. Prepare the knowledge
We clean and structure the agreed content, track sources and versions, and flag decisions your team needs to make.
-
3. Validate and hand off
We run the agreed checks and deliver the prepared files, handoff guide, findings, and maintenance guidance.
Prepared by Richard Gregorio
Richard is an AI technical architect with more than twenty years in enterprise technology, including multi-national Salesforce implementations. Greg Tech Advisory applies that experience to the source content an assistant retrieves.
Richard's professional experience
- AT&T
- IBM
- Salesforce
- USAA
- Tata Consultancy Services
- Bradesco
- C&A
- BTG
- Cielo
- Yduqs
- RX Global
- Minsait
- CloudKitchens
Questions
Does this include building the AI assistant?
No. The sprint prepares the agreed knowledge collection. Assistant development, integrations, and production deployment are scoped separately.
Will this guarantee accurate AI answers?
No. We address source-content issues and validate against agreed checks. Answer quality also depends on retrieval, model behavior, and application configuration.
What determines the final price?
Content volume, formats, extraction complexity, and validation requirements. Scope and price are agreed before kickoff.
What happens when sources disagree?
We document the conflict and ask your designated business owner to confirm the authoritative information.
Do you upload into our vector store or knowledge index?
Not in the baseline sprint. You receive the prepared files, manifest or front-matter rules, and step-by-step loading guidance. Production ingestion, indexing, and connectors are separately scoped if you need hands-on deployment.
When does the three-week window start?
After scope approval and receipt of all required inputs and access. Outstanding content decisions may affect completion.
Check Your Project Fit
Tell us about your assistant and its source content so we can assess whether the sprint fits your project.